AI Reads Ultrasound and MRI Together to Predict Breast Cancer Subtypes Without a Biopsy

Every breast cancer is not one disease. Under the microscope, tumors split into molecular families—hormone receptor-positive luminal cancers, HER2-driven cancers, and triple-negative tumors—each with its own biology, its own preferred treatment, and its own prognosis. Getting that classification right before therapy begins can determine whether a patient receives endocrine therapy, targeted drugs, or chemotherapy first. Now, a large two-center study from China suggests that artificial intelligence can read the fingerprints of these subtypes directly from routine medical images, potentially giving clinicians a non-invasive preview of tumor biology before a single needle is inserted.

The research, published in Holistic Integrative Oncology, was led by a team at Sun Yat-sen University Cancer Center in Guangzhou. The investigators built what they call a combined radiomics model, one that fuses quantitative features extracted from two of the most common imaging tests in breast care: ultrasound and dynamic contrast-enhanced magnetic resonance imaging, or DCE MRI. Their goal was ambitious but precise—to distinguish luminal from non-luminal breast cancers preoperatively, using only images that patients already receive in standard clinical workups.

The clinical logic behind the effort is straightforward. Luminal cancers, which express hormone receptors, typically respond well to endocrine therapy and carry the highest five-year survival rates. Non-luminal tumors—HER2-positive and triple-negative cancers—behave differently. Because they tend to be more sensitive to chemotherapy, guidelines often recommend neoadjuvant treatment given before surgery for these patients. Knowing the subtype in advance therefore shapes the entire treatment sequence. Today, that knowledge comes from a core needle biopsy analyzed with immunohistochemistry, an invasive procedure whose results depend on where in the tumor the sample was taken and how a pathologist interprets the staining.

That sampling problem is not trivial. Studies cited by the team indicate that roughly 5.9 to 18.5 percent of biopsy specimens fail to match the molecular subtype of the final surgical specimen, a discrepancy rooted in tumor heterogeneity—different regions of the same cancer can carry different receptor profiles. A non-invasive imaging-based prediction, the researchers argue, could complement biopsy by assessing the whole lesion rather than a sliver of it, and could serve as a decision-support tool when biopsy is delayed, contraindicated, or declined.

To build the model, the team assembled 655 patients with primary invasive breast cancer from two hospitals. A cohort of 551 patients from Sun Yat-sen University Cancer Center was split in a 3:1 ratio into a training set of 413 cases and an internal test set of 138 cases, while 104 patients from the First Affiliated Hospital of Gannan Medical University formed an independent external validation set—an essential step, since many previous AI models in this field were trained and tested on data from a single center, inflating their apparent accuracy. Importantly, images were acquired on many different scanners from multiple manufacturers, with varied acquisition parameters, testing whether the model could generalize beyond any single device.

The technical pipeline was exhaustive. For each patient, radiologists delineated the tumor on a representative ultrasound image and, slice by slice, on the DCE MRI phase showing the strongest enhancement. Software then extracted thousands of quantitative features—1,561 from each ultrasound image and 1,316 from each MRI scan—capturing first-order statistics, lesion shape, and intricate texture patterns, including features derived from wavelet and other image transformations. Through a multi-stage filtering process involving inter-observer agreement checks, statistical testing, correlation pruning, and LASSO regression with tenfold cross-validation, the combined signature was distilled to just 25 key features: nine first-order, one shape, and fifteen texture descriptors.

Those features were fed into eight different machine learning algorithms, from logistic regression and random forests to gradient boosting methods and support vector machines. Because non-luminal cancers made up only about a quarter or less of each dataset, the team applied SMOTE, a synthetic over-sampling technique that generates new examples of the minority class to prevent the algorithms from simply defaulting to the more common luminal category. The effect was striking in some cases: the support vector machine’s performance on the internal test set jumped from an AUC of 0.731 without balancing to 0.872 with it, although the authors note that this sensitivity to preprocessing should be kept in mind when reproducing the model.

The support vector machine ultimately emerged as the winner. Its combined model achieved an AUC of 0.973 in the training set and 0.872 in the internal test set, significantly outperforming the ultrasound-only model (AUCs of 0.701 and 0.751) and the MRI-only model (0.919 and 0.742). On the external validation cohort from the second hospital, the combined model posted an AUC of 0.835, again numerically ahead of the single-modality models at 0.684 for ultrasound and 0.740 for MRI, though the differences there did not reach statistical significance. Perhaps most clinically meaningful were the model’s safety-oriented metrics: a sensitivity of 0.833 and a negative predictive value of 0.927 in external validation, and even higher values of 0.926 and 0.976 in the internal test set. In plain terms, the model rarely missed a non-luminal cancer—the very patients who need neoadjuvant chemotherapy—while patients the model classified as negative were very likely to have luminal disease suitable for endocrine therapy.

Decision curve analysis reinforced the statistical results, showing that the combined model delivered the highest net benefit across a range of clinical threshold probabilities compared with either single-modality approach. The authors position the tool not as a replacement for pathology but as a complementary layer in the diagnostic workflow: a preoperative risk-stratification aid that could help flag patients likely to benefit from chemotherapy before surgery, or provide guidance in settings where tissue sampling is impractical. Because the model draws on two-dimensional ultrasound and three-dimensional MRI views simultaneously, it captures complementary information—ultrasound excels at lesion margins and echotexture, while contrast-enhanced MRI reveals how a tumor perfuses and enhances, a window into its proliferative activity.

The study is not without caveats, which the authors acknowledge candidly. It was retrospective, meaning image quality and completeness could not be fully controlled, and the numbers of HER2-positive and triple-negative cases were relatively small, reflecting their real-world prevalence. Manual segmentation of MRI lesions, performed layer by layer, introduces subjectivity and labor costs that future deep learning approaches could automate. The team calls for prospective, multi-center collection of more cases to sharpen sensitivity and positive predictive value further. Still, the demonstration that a radiomics model trained on heterogeneous, multi-vendor images from two hospitals can hold its performance on completely unseen external data marks a meaningful step toward bringing AI-assisted molecular subtype prediction out of the research lab and into the breast clinic—where a few minutes of computation on images patients already have could help ensure the right therapy reaches the right tumor at the right time.

Subject of Research: Radiomics-based prediction of breast cancer molecular subtypes using ultrasound and DCE MRI

Article Title: The radiomics models based on ultrasound and DCE MRI for predicting molecular subtypes of breast cancer

Article References: Mao, R., Zheng, H., Tang, X., Yang, L., Zhang, Y., & Zhou, J. (2026). The radiomics models based on ultrasound and DCE MRI for predicting molecular subtypes of breast cancer. Holistic Integrative Oncology, 5(1), Article 28. https://doi.org/10.1007/s44178-026-00243-2

Image Credits: AI Generated

DOI: 10.1007/s44178-026-00243-2

Keywords: breast cancer, radiomics, molecular subtypes, ultrasound, DCE MRI, machine learning, support vector machine, neoadjuvant therapy, luminal cancer, triple-negative breast cancer, non-invasive diagnosis, external validation